LlamaIndex launches ExtractBench benchmark and new Agentic Plus tier for LlamaExtract

LlamaIndex announced ExtractBench, an open, public benchmark designed to evaluate document extraction from enterprise paperwork. LlamaIndex states that the benchmark comprises 370 documents (4,869 pages) across eight business domains and 67 document types, and measures long‑record completeness, handling of scans and handwriting, grounding, and extraction cost. The dataset, evaluation harness, and methodology are released on HuggingFace and GitHub, and a companion paper is posted on arXiv. In the initial release, LlamaIndex evaluated 14 systems—including frontier vision‑language models, coding agents, and specialized extraction APIs—using ExtractBench. LlamaIndex reports that LlamaExtract’s new Agentic Plus tier achieved the highest overall accuracy, with a 95.6 % value‑F1 score and leading grounding scores, at a cost of 8.1 ¢ per page. By comparison, the other LlamaExtract tiers (Agentic and Cost‑Effective) scored lower, and many VLM‑based systems fell sharply on long documents. LlamaIndex describes ExtractBench as "the most comprehensive extraction benchmark" and positions Agentic Plus as its top‑performing offering for production‑scale document extraction. The company invites users to try the benchmark with 10,000 free credits and to join a live webinar on August 26. All benchmark data and code are publicly available, enabling independent replication of the results.